title: A systematic machine learning approach to quantifying coverage and representation bias in population estimates from mobile phone app data author: - name: Carmen Cabrera affiliation: “1” - name: Francisco Rowe affiliation: “1” address: - code: “1” address: Geographic Data Science Lab, Department of Geography and Planning, University of Liverpool, Liverpool, United Kingdom. - code: “” address: corresp_author_name: “Carmen Cabrera” corresp_author_email: “C.Cabrera@liverpool.ac.uk” subject: - “Mobile phone data” - “Human mobility” - “Explainable AI” - “Spatial analysis” keywords: - Mobile phone data - Human mobility - Location - Explainable AI - Spatial analysis abstract: | Traditional data sources such as censuses and surveys are costly, infrequent, and often unavailable in crisis-affected regions. User location data derived from GPS-enabled mobile phone (MP) applications offer near-real-time, high-resolution insights into population distribution, but unequal access to and use of mobile technologies introduces biases that threaten representativeness. Existing bias assessments typically require demographic attributes, which are often unavailable, and focus on national-level estimates. We present a generalisable framework to measure and explain biases in aggregated MP app data without the need for individual-level demographic data. The framework quantifies coverage bias, which relates to the share of the population captured, at national and subnational levels, evaluates spatial heterogeneity and clustering, and models the geographic variation of bias as a function of area-based covariates using explainable machine learning. We illustrate the framework using four MP app datasets for the UK, aligned with the 2021 national census. We find that MP data consistently achieve higher coverage than major national surveys, though bias varies spatially and by data source. Multi-application datasets generally reduce (?) coverage bias relative to single-application sources. X emerges as a consistent and important factor in determining the local magnitude of coverage bias. ethics: | Please provide details on the ethics.

data_accessibility: | Please provide details on the data availability. author_contributions: | Please provide details of author contributions here. conflict_of_interest: | Please declare any conflict of interest here. funding: | Please provide details on funding disclaimer: | Please provide disclaimer text here. acknowledgements: | Please include your acknowledgments here, set in a single paragraph. Please do not include any acknowledgments in the Supporting Information, or anywhere else in the manuscript. bibliography: sample.bib lineno: false site: bookdown::bookdown_site output: bookdown::pdf_book: base_format: rticles::rsos_article editor_options: markdown: wrap: 72 —

The main document should include:

Title (no more than 150 characters)

Author names and affiliations, and ORCID iDs where available

Abstract (no more than 200 words). (This will be used in reviewer invitation emails so please think about how to describe your work to make it easy for a potential reviewer to determine whether they would be suitable.)

All main manuscript text. Ensure that all figures, tables and any relevant supplementary materials are mentioned within the text References

Acknowledgements and funding statement (ensure that you have included grant numbers and the names of any funding providers)

Tables with captions Figure captions with credits where relevant

Introduction

Traditional data streams, such as the census and surveys have been the primary official source to provide a comprehensive representation of national populations in countries worldwide. However, fast-paced societal changes and emergency disasters, such as climate-induced hazards and COVID-19 have tested and accentuated weaknesses in traditional data systems [@green2021]. Traditional data systems often provide data in infrequent and coarse temporal and geographical resolutions [@rowe23-bigdata]. Generally they are expensive to maintain and operate, and are slow taking months or years since they data are collected to their release [@rowe23-bigdata]. Data collection from climate- or conflict-impacted areas is generally unfeasible because of restrictions due to high levels of insecurity and risk [@iradukunda2025]. Yet, fast-paced societal changes require high frequency, granular and up-to-date information to support real-time planning, policy and decision making.

At the same time, we have seen the confluence of two diverging trends in data availability. On the one hand, growing evidence of declining survey response rates across many countries over the last 20 years is accumulating [REF]. Dwindling numbers in surveys can represent distorted picture of society [REF]. On the other hand, significant advances in sensor technology, computational power, storage and digital network platforms have unleashed a data revolution producing large trails of digital trace data [REF]. These data are now routinely collected and stored. They offer spatially granular, frequent and instant information to capture and understand human activities at unprecedentedly high resolution and scale, with the potential to produce real-time actionable intelligence to support decision making [REF]. Hence, national statistical offices are actively seeking to integrate these data into their national data infrastructure [REF].

Mobile phone data (MPD) collected via GPS- and IP-based technology have become a prominent source of nontraditional data to monitor population changes. Increasing usage of mobile services on smartphones and wearable devices have resulted in the generation of large volumes of geospatial data, offering novel opportunities to advance understanding of spatial human behaviour, and thus revolutionise research, business and government decision making and practices [@rowe23-bigdata]. MPD are now a core component of the digital economy, creating new market opportunities for data intelligence businesses, such as Cuebiq/Spectus, Safegraph and Locomizer. They have been used to create critical evidence to support policy making, prominently during the COVID-19 pandemic. In research, MPD have been used to develop innovative approach to infer mode of transport [REF], monitor footfall changes [REF], profile daily mobility signatures [REF], sense land use patterns [REF], predict socioeconomic levels [REF], define urban extents [REF], quantify tourism activity [REF] and estimate migration and population displacement [REF].

However, the use of MPD present major epistemological, methodological and ethical challenges [@rowe23-bigdata]. A key unresolved challenge is potential biases in MPAD compromising their statistical representativeness and perpetuate social injustice [REF]. Biases reflect societal digital and socioeconomic inequalities. Biases emerge from differences in the access and use of the mobile phone applications used to collect MPD [@wesolowski13-biases]. Only a fraction of the population in a geographical area owns a smartphone, and even an smaller share actively uses a specific mobile phone app. In the UK, for example, 98% of the adult population have a mobile phone and 92% of this population use a smartphone [@ofcom23], but a smaller percentage actively use Facebook (70%) or Twitter (23%) [@statista24]. Additionally, biases emerge from differences in the access and use of digital technology across population subgroups reflecting socioeconomic and demographic disparities. For instance, wealthy, young and urban populations generally have greater access and more intensively use of mobile phone applications, and therefore tend to be over-represented in MPD [REF].

The use of biased MPD can thus have major practical and societal implications. If used uncorrected, MPD reproduce selective patterns of smartphone ownership and application usage, rendering inaccurate or distorted representations of human population activity. Such representations disproportionately reflect behaviours of younger, urban and higher-income users while underrepresenting marginalised or less-connected groups. Distorted representations based on biased MPD can thus misguide decision making, policy and planning interventions, and thus amplify existing socio-economic disparities. In practice, existing applications of MPD often use uncorrected population statistics derived from MPD and have thus been constrained to offer a partial picture for a limited segment of the overall population. Such data can only afford to provide rough signals about the spatial distribution of (e.g. spatial concentration), trends (e.g. increasing) and changes (e.g. low to high) in populations [@rowe22-sensing-ukraine]. They have cannot provide a full representation of the overall population.

Efforts have been made to measure and assess biases in aggregate population counts from digital data sources. Existing analyses typically measure the extent of bias measuring the system-wide difference in the representation of population counts from digital platforms and censuses. To estimate the representation of digital data sources, the penetration rate is computed as the active user base of a digital platform over the census resident population. Existing analyses have thus been able to established systematic gender, age and socio-economic biases in population data obtained via API (or Application Programming Interface) from social media platforms, such as Facebook and Twitter/X. However, this approach requires information on the demographic and socio-economic attributes of the collected sample and has focused on estimating biases at the country level. Yet, these attributes are generally unavailable for MPD, and biases may vary widely across subnational areas. What is missing is an systematic approach to measure biases in population counts from digital platforms, when population attributes are unknown, and quantify the geographic variability in the extent of biases in these data.

To address this gap, this paper aims to establish a standardised approach to empirically measure the extent of biases in population data derived from digital platforms, and identify their key underlying contextual factors across subnational areas. We seek to address the following research questions:

What is the comparative extent of population coverage of digital sources relative to widely-used traditional surveys?
How systematic is the association between larger biases and the over-representation of rural, more deprived, child and elderly populations?
To what extent, are population data assembled from multiple applications versus single applications associated with lower bias?

Our approach proposes a statistical indicator of population coverage to measure the extent of bias, and uses explainable machine learning to identify key contextual factors contributing to spatial variations in the extent of bias. Biases in digital trace data can emerge from multiple sources, such as algorithmic changes, device duplication and geographic location accuracy [REF]. We do not intend to identify these individual sources of error. We focus on quantifying the extent of ``cumulative’’ bias; that is, the resulting bias from the accumulation of these error sources. We use data collected from single and multiple mobile phone apps, and compare their results. As outlined above, we test the extent to which biases can be mitigated by leveraging information from multiple apps encompassing a more diverse user population. Specifically, we use two single-app (i.e. Facebook and Twitter/X) and two multi-app providers (i.e. Locomizer and a European provider). We focus on the use of aggregated population counts as this has become a common ethical and privacy-preserving practice for companies to provide access to highly sensitive data for social good.

Our study makes two key contributions.

Methodological contribution i.e. what we hope to achieve with our approach / quality assessment framework ideas + start setting standards of good practice in the use of MPD.
Substantive contribution - systematic evidence identifying key predictor of biases + do we find evidence of lower biases / greater population coverage for multi-app better than single app?

Data

We propose a systematic framework to measure and explain biases in population count data derived from mobile phones (MPs). We use four datasets to illustrate this framework, collected in or around March 2021 to align as closely as possible with the dates of the most recent census in the area of study, hence enabling temporally consistent comparisons. We focus on aggregated population counts, which are commonly used in mobility research, as a privacy-preserving and ethically responsible data format. The datasets include sources derived from a single MP application (Meta and Twitter/X) as well as from multiple MP applications, each capturing distinct user groups through different data generation mechanisms. These differences allow us to assess how source characteristics influence population coverage and representativeness. The multi-application sources are referred to as Multi-app1, whose provider name cannot be disclosed due to a non-disclosure agreement, and Multi-app2, provided in its raw format by the company Locomizer. Table 1 summarises the main characteristics of each dataset, including the source type, form of data collections, temporal granularity, temporal coverage, spatial resolution, access method and data acquisition cost. Further details of access and processing for each data source are provided in the following subsections.

It is important to note that, while Twitter/X is not exclusively accessed via mobile devices and its location data are not always collected via GPS, it has nonetheless been widely used in population and mobility research for its ability to capture patterns at high spatio-temporal resolution and across broad geographic areas. Additionally, the Twitter Academic API is no longer available for free data collection, limiting access to new data. Despite these limitations, we include Twitter/X in our analysis as a representative single-application data source. Archived datasets, such as the one used in this study or the Harvard Geotweet Archive (https://gis.harvard.edu/data) continue to support population and mobility research.

Table 1. Summary description of data sources derived from mobile phone apps used in the article.

Twitter

We use an anonymised, analysis-ready dataset of active X (previously Twitter) accounts in the UK, originally collected via the Twitter Academic API. Like in the data for Meta, we take the number of unique accounts as a proxy for the number of unique users. The data consists of monthly records for the location of unique Twitter accounts, spatially aggregated across the UK, and is openly available at https://github.com/c-zhong-ucl-ac-uk/Twitter-Internal-Migration. Geolocation is obtained either directly from geotagged tweets or through manual geocoding using bounding boxes provided by the API, based on the IP address of the posting device (for methodological details, see [@wang2022]). The full dataset includes approximately 161 million tweets from February 2019 to December 2021. For this study, we restrict the analysis to March 2021 to align with the timing of the 2021 UK Census, during which 125,637 user home locations were identified. Home locations are assigned to Local Authority Districts (LADs) using a frequency-based detection algorithm, further described in [@wang2022].

Multi-app1

We sourced data from a location analytics company that collects GPS data from approximately 26% of smartphones in the UK. The raw data consist of anonymised device-level GPS traces collected via a range of smartphone applications, where users have explicitly granted location-sharing permissions. We consider the number of devices as a proxy for the number of unique users, although it could be the case that some users have more than one device. The dataset spans a 7-day period corresponding to the first week of April 2021 and includes 443,553,155 GPS records. Although the dataset does not perfectly align with the official 2021 UK Census date, the temporal proximity ensures a high degree of comparability.

To infer the place of residence of users, we apply a commonly used rule-based classification method, following approaches outlined in [@wang2022; @zhong24working]. Specifically, the place of residence associated with a device is defined as the location with the highest number of GPS records recorded during nighttime hours (10 PM–6 AM). To be classified as a residence, a location must account for more than 50% of the device nighttime records. Furthermore, the number of nighttime records during the observation period must be at least 2. For comparability across data sources, all identified residence locations are aggregated to the level of Local Authority Districts (LADs). Using this method, we detect 1,536,922 home locations.

Multi-app2

Our analysis includes a second source of analysis-ready location data, which is openly-available on GitHub (https://t.ly/dzlzB)). This dataset has already been processed to identify the home location of users according to the methodology described in [@zhong24working]. The raw data is collected by a UK-based data service company, which licenses mobile GPS data from 200 smartphone apps and applies pre-processing methods to ensure user privacy and anonymity. The dataset covers the entire UK for November 2021 and includes inferred home and work locations for 630,946 users.

While this period does not exactly coincide with the 2021 UK Census, the difference of less than a year is considered sufficiently close for our analysis. To ensure consistency across datasets, we further process the data by aggregating it spatially from the Middle Layer Super Output Area (MSOA) level to the Local Authority District Level (LAD).

Census data

In addition to the mobile phone app data sources described above, we use resident population counts from the 2021 UK Census, aggregated at the LAD level. UK Census is trusted as the official benchmark for population counts, so we take these as the ground truth for comparing estimates derived from each digital dataset. We also draw on a set of area-based covariates from the census, covering demographic, socioeconomic, and housing characteristics, along with the geographic coordinates of each LAD centroid. These variables are used to investigate and explain the contextual factors most strongly associated with the magnitude and spatial variation of bias in the digital trace data. The full list of covariates is provided in Table 2.

Table 2. List of covariates used in the analysis, grouped by thematic category. Variables are derived from the 2021 UK Census and aggregated to the Local Authority District (LAD) level, except for geographic coordinates, which correspond to the centroid of each LAD.

Methods

We introduce a framework to measure and explain biases systematically in population count data derived from mobile phones (MPs). This framework consists of three stages. In the first stage, we introduce a metric to quantify bias related to population coverage within a given geographic area. In the second stage, we calculate the coverage metric for different data sources and subnational geographic areas. We then analyse its variability and assess whether it is evenly distributed across geographies. Detecting such patterns is important for understanding the limits of data applicability, and to assess whether spatial dependencies should be considered in the third stage of the analysis. In the third stage, we use explainable machine learning to model the variation in coverage bias as a function of demographic and socioeconomic covariates derived from the 2021 UK census. This modelling approach allows us to model the magnitude of bias across areas and, importantly, to quantify the relative importance of each covariate, so it is possible to identify the population characteristics (e.g. age structure, income levels, educational attainment) that are most strongly associated with overrepresentation or underrepresentation in the different sources of MP app data. Figure 1 provides an overview of the methodological workflow, which includes data acquisition, bias measurement, comparative analysis with national surveys, spatial analysis, and bias explanation through modelling.

Measuring coverage bias

We define a metric to quantify the magnitude of coverage bias in each subnational area. This metric is based on the population coverage of the dataset, which we compute as the ratio of the population captured by dataset \(D\) (sample size) in a geographic area \(i\), denoted as \(P_i^D\), to the total local population of the same area, \(P_i\). Formally, the coverage \(c_i\) for area \(i\) is given by: \[\begin{equation} c_i = \dfrac{P_i^D}{P_i} \times 100. \end{equation}\] The resulting ratio \(c_i\) is assumed to take values between \(0\) and \(100\), with 100 representing full population coverage. If users have multiple accounts, the ratio can exceed \(100\), since the total sample size could be greater than the local population of area \(i\).

We then define the size of bias \(e_i\) for area \(i\) as:

\[\begin{equation} \label{eq:size-bias} e_i = 100 - c_i \end{equation}\]

A value of \(e_i = 0\) indicates a lack of coverage bias, which corresponds to full population coverage (\(c_i = 100\)). We use this bias indicator to analyse the magnitude and distribution of coverage bias across multiple sources of data and geographic areas.

Identifying spatial patterns of bias

For each data source, we compute the coverage bias metric at the subnational level and examine its geographic variation. This stage has two main objectives. First, to assess whether bias is evenly distributed across geographies. Second, to determine whether spatial effects are sufficiently strong to consider them in the subsequent methodological stage through the inclusion of spatial lag terms in the explainable machine learning model.

To evaluate the variability of bias across geographies, we first conduct exploratory analyses using thematic maps and histograms. To formally test for spatial clustering, we calculate Moran’s I statistic for each dataset. Because Moran’s I is sensitive to the definition of spatial relationships, we evaluate four alternative spatial weighting schemes: 1) queen neighbourhood, 2) k-nearest neighbours, 3) distance band (set by algorithm), and 4) distance band (set by user) [REF]. Comparing results across these schemes enables us to assess the robustness of clustering patterns to different neighbourhood definitions. In the main body of the paper, we report Moran’s I values obtained using scheme 1, as it produces the highest statistic across datasets when statistically significant, thereby providing the most conservative test for the presence of spatial clustering. Results for the other schemes are provided in the Supplementary Information.

Finally, to examine whether bias is associated with population size, we generate scatterplots comparing population counts from the digital data sources with census population counts. If bias varies with population size, we would expect systematic departures from proportionality. Conversely, an approximately linear relationship through the origin with a stable slope would suggest that bias is largely independent of population size. We quantify this relationship by computing Pearson’s correlation coefficient.

Explainable machine learning

We used explainable machine learning to identify the key predictors of population bias and how these the importance of these predictors varies across geographical areas. Existing evidence based on social media suggests that population location data from digital platforms are biased over-representing urban, wealthy and young-adult populations [REF]. We therefore modelled our measure of population bias from Equation~\(\ref{eq:size-bias}\) as a function of key area-level attributes reflecting geographical differences in engagement and access to digital technology across demographic, socioeconomic, household, housing and location factors. Table 2 reports the set of predictors included in our analysis. We used data from the 2021 census for England and Wales to measure these predictors.

We used an eXtreme Gradient Boosting (XGBoost) algorithm. XGBoost is an ensemble that combines outputs from multiple models to produce a single prediction and represents an efficient and scalable adaptation of the gradient boosting machine algorithm proposed by [@friedman2001a]. It utilises gradient descent to improve model performance, and decision trees are built iteratively, with each tree built to minimise the error residuals of a preceding iteration. XGBoost has been optimised for scalability and computational efficiency, providing high predictive accuracy with limited training time [@chen2016; @nielsen2016tree]. XGBoost has also become one of the most widely-used off-the-shelf machine learning models in applied settings because of its built-in regularization that mitigates overfitting, sparsity-aware tree construction and parallelisation efficiency [@chen2016]. It can accommodate nonlinearities and is robust to multicollinearity [@chen2016]. We fitted the following XGBoost regression model.

\[\begin{equation} \label{eq:xgb-model} \widehat{e}_i = \sum_{m=1}^M f_m\bigl(D_i, S_i, H_i, U_i, L_i\bigr), \quad f_m \in \mathcal{F} \end{equation}\]

\(e_i\) is our measure of population bias. \(f_m\) denotes an individual regression tree from the boosted ensemble \(\mathcal{F}\) and \(M\) is the total number of trees. The input variables \(D\), \(S\), \(H\), \(U\), \(L\) represent key demographic, socioeconomic, housing, household, and locational attributes of area \(i\), respectively. The model iteratively learns the contribution of each feature to the prediction of the bias indicator \(e_i\), allowing for complex, nonlinear interactions.

To implement Equation~\(\ref{eq:xgb-model}\), we randomly split the data into training (80%) and testing (20%) sets to ensure robust model evaluation. We used 10-fold cross validation to train models and performed grid search over learning rates, tree depths, subsample ratios, and regularisation penalties to identify optimal hyperparameters. We applied regularisation penalties including L1 (Lasso) and L2 (Ridge) terms to penalise overly complex trees, promote feature sparsity, improve model generalisation and mitigate multicollinearity among predictors. XGBoost’s tree-based structure additionally handles multicollinearity by hierarchically selecting the most informative splits [@chen2016]. We then fitted a final model on the full training set using these tuned settings of optimal parameters and evaluated on the held-out test set. We evaluated models based on the number of trees minimising the root mean squared error (RMSE), the convergence of training and test error, and difference between predicted and observed values.

Results

In this section, we illustrate our proposed methodological framework on four sources of digital data derived from mobile phone (MP) apps. As described in the Data section, these sources include data for the UK, collected in or around March 2021 to align as closely as possible with the reference date of the most recent national census. This temporal alignment ensures comparability, as the census data serve as the ground truth against which population counts from the digital sources are evaluated.

The extent of bias varies across data sources

We begin by quantifying coverage bias in each of the mobile phone (MP) datasets, defined as a function of the proportion of the total population captured by the dataset, as defined in Equation \(\ref{eq:size-bias}\). Coverage bias is first computed at the national level for each MP data source. To contextualise these results, we compare the MP sources with several widely used traditional datasets, including major UK surveys available through the UK Data Service [@ukdataserviceSurveysData]. Figure X presents these comparisons across data sources, showing on the top x-axis the population coverage (expressed as the number of respondents or subjects per 1,000 people) and on the bottom x-axis the corresponding measure of coverage bias.

Size of coverage bias (bottom x-axis) and population coverage per 1,000 population (top-x-axis), by data source.

Figure \(\ref{fig:survey}\) suggest that, with comparatively greater population coverage and lower coverage bias, MP data have a strong potential to support large-scale empirical analyses. However, high population coverage alone does not ensure the data is representative of different population groups. In surveys, specific strategies are usually implemented during the data generation process to improve the statistical representativeness of the sample. For example, sampling techniques such as stratified sampling or cluster sampling can be applied so that the sample reflects the broader population of interest. After sampling, if certain groups remain under-represented, responses can be adjusted using post-stratification techniques [@lohr2021]. However, even when these strategies are applied, there is no guarantee that the survey will be fully representative of the broader population of interest [@cochran1977sampling]. This is because representativeness can only be achieved with respect to a finite set of attributes (e.g. age, gender, income levels, location, etc.). Ensuring perfect representativeness would only be possible by surveying the whole population, which is practically not feasible.

With MP data, achieving statistical representativeness is even more challenging. Unlike survey data, which is actively collected using structured sampling methods, MP data is generated passively as a byproduct of digital interactions, transactions, or device usage, without any control over who is included in the dataset. Furthermore, by the time this data reaches researchers or analysts, it is often anonymised, and does not contain demographic identifiers. As a result, it is not possible to apply the standard post-stratification weighting techniques that are typically used to adjust survey or census data for improved representativeness.

We argue that, even though we do not always have specific demographic information of the individuals captured through digital trace data, we can infer some of these characteristics by leveraging the spatio-temporal granularity of MP data. This is a necessary first step to understand which population groups might be overrepresented or underrepresented in different sources of MP data. This information is necessary to develop subsequent data adjustment strategies that can improve the representativeness of the data relative to the target population.

The spatial variability of bias

We next leverage the fine-grained geographic resolution of the MP app data sources to examine coverage bias at subnational levels. Analysing the distribution of coverage bias across geographies allows us to identify uneven patterns of population coverage across areas and assess whether these patterns exhibit spatial clustering. On the one hand, these assessments can help evaluate the limitations in the applicability of each dataset for further research, and on the other, they inform whether spatial effects should be incorporated into the subsequent modelling stage.

Figure \(\ref{fig:size-bias-spatial}\) shows multiple visualisations that reflect the distribution of coverage bias at the Local Authority District (LAD) level. Each row corresponds to a data source and includes three elements. First, a hexagonal cartogram for the size of coverage bias in each LAD, representing the LADs as hexagons of equal size to aid comparability of LADs while maintaining their relative positions. The maps are presented alongside the associated Moran’s I and corresponding p-value as a measure of spatial autocorrelation. Second, a histogram showing the statistical distribution of coverage bias across LADs. Third, a scatter plot comparing the population counts derived from each of the MP app data sources and the corresponding census population for each LAD. The scatter plot is accompanied by the Pearson correlation coefficient and corresponding p-value.

Our spatial analysis reveals several patterns in the distribution of coverage bias which are consistent across the mobile phone (MP) datasets. First, all data sources exhibit noticeable geographic variability in coverage bias, as evidenced by the spread of values in their respective distributions, as displayed in the histograms. This variability highlights that for a given data source, the degree of representation can differ substantially between areas.

Second, despite the variability in coverage bias, the maps do not reveal strong geographic clustering patterns. Bias values fluctuate across longitude and latitude, and there are no clear north–south or east–west gradients. This observation is quantitatively supported by the Moran’s I statistics, which are generally statistically significant, but close to zero. Their small magnitude indicates that spatial clustering is weak at the LAD scale. Consequently, we conclude that it is unnecessary to include spatial dependence terms (e.g. spatial lag) in the subsequent modelling stage of our framework.

Third, the variability in coverage bias is not explained by the absolute population size of each subnational unit. Scatter plots comparing population counts derived from MP app sources and from the census reveal strong linear relationships. This is quantitatively supported by Pearson correlation coefficients consistently close to one and statistically significant.

These findings suggest that coverage bias at the LAD level is not explained by geographic location or population size, but rather, by other area-level characteristics such as demographic or socioeconomic composition. We examin these factors in the third stage of our analysis.